The Cost of a Wrong Agent Action, Measured in Hours Not Dollars

The pitch for an AI agent is always in hours saved. The argument against it is always in vague talk of risk. Those two cannot be compared, so the loudest person wins. Measure the cost of a wrong agent action in hours to correct, and both sides of the decision suddenly use the same unit.

The method is simple. For each thing an agent might get wrong, estimate how much staff time it takes to notice, undo and repair. Then compare that total against the time the agent saves.

Why hours beat money here

Money looks precise and usually is not. Converting a mistake into a figure means guessing at reputational damage, opportunity cost and an hourly rate you probably do not have. Every one of those guesses is arguable, so the conversation becomes an argument about assumptions.

Hours are countable. People can tell you how long it took to fix something last time. Nobody argues about whether it took three hours or three days, because somebody remembers.

Hours also make the two sides commensurable, provided the savings side is counted with the same care and you are measuring time saved in a way that survives scrutiny. If a job saves four hours a week and its failures cost twelve hours a quarter to repair, that is a decision anyone in the room can make.

The four parts of a correction

Total correction time is bigger than the fix itself. Break it into four:

• Detection: the time between the mistake happening and somebody noticing. This is the part everyone forgets, and it is often the largest.
• Investigation: working out what happened, what else was affected and whether it is still happening.
• Repair: the actual undo. Reissuing the invoice, correcting the record, sending the follow up message.
• Aftermath: the meeting, the note to the client, the process change, the extra checking everyone does for a fortnight.

Aftermath is the one that scales with visibility. A wrong internal summary has almost none. A wrong message to a customer has a lot.

The cost of a mistake is not the time to fix it. It is detection plus investigation plus repair plus aftermath, and for anything a customer sees, aftermath is usually the biggest of the four.

Building the estimate

You do not need a model. You need one hour with the people who do the work.

For each candidate job, write down the two or three ways it realistically goes wrong. Not the exotic failures, the ordinary ones. Then ask the team to estimate the four parts for each, based on times it has already happened with a human doing the job.

An example shape, using invented numbers purely to show the arithmetic rather than to claim anything about your business: if a wrong internal status summary costs roughly an hour end to end and a wrong customer invoice costs most of a day, those two jobs deserve very different levels of review even though both look like "generate a document".

Do this for three or four jobs and the ranking usually becomes obvious without anyone having to argue.

Multiply by how often, not by how badly

A rare catastrophe and a frequent nuisance can carry the same annual cost. Once you have correction hours per incident, estimate how often the mistake happens. Error rate times correction hours gives you an expected cost per month in the same unit as your savings.

This is also where detection time earns its place. A failure caught in five minutes and a failure caught in five weeks may need identical repair work, but the second one has spread. Anything with slow detection deserves a lower tolerance for error.

What the numbers change

Three decisions usually flip once you have hours on both sides.

Which job goes first. Teams tend to automate the most annoying task. The arithmetic often points instead at a duller task with faster detection and cheaper repair.

How heavy the review should be. A job whose failures cost twenty minutes does not need the same approval process as one whose failures cost two days. Uniform review is either too slow everywhere or too thin where it counts.

Whether to automate at all. Some jobs save less time than their mistakes cost. That is a reasonable answer, and it is much easier to say out loud with a number behind it.

Keep the measurement going

The estimate is a starting point, not a verdict. Watch what you are counting as well as how much, because a pilot can look healthy while measuring the wrong thing. Once an agent is live, record every correction and how long it took. After a quarter you can replace the guesses with observations.

Log the near misses too. An output a reviewer caught and rewrote is a mistake that did not reach the world, and counting those tells you whether your review step is doing work or sitting idle. Reviewing is itself a cost in hours, and it is the one nobody budgets for.

Start with one job, one hour of estimating and four columns. It is a cheaper way to settle the automation argument than another round of opinions.